Artificial Neural Networks as a quality loss function for Six Sigma

In this study, Artificial Neural Networks (ANNs) are proposed to be used as a quality loss function for Six Sigma projects. An industrial data set consisting of power consumption rates of refrigerators and thermal camera readings around their compressors is analysed. For industrial data, relationships between inputs and outputs can be nonlinear and complex. Therefore, traditional statistical models may result in poor inferences. At this point, ANNs emerge as effective tools because of their ability to learn nonlinear and complex relationships. While Six Sigma remains as the major quality initiative, its popularity started to decline among industrial practitioners. Since the methodology was established, Six Sigma toolbox was not radically improved. Therefore, to enhance Six Sigma toolbox, an ANN-based structure is proposed to detect the refrigerators not complying with power specifications through thermal camera readings. Four quality loss function models are compared: one-dimensional parabolic Taguchi loss functions, multivariate Maximum Likelihood cost, logistic regression and finally ANNs. The analyses are conducted by Monte Carlo cross-validation to obtain precision-recall curves for these methods. The ANN-based cost function is shown to outperform other three methods. Finally, ANNs are found to be an effective tool which may bring new dimensions to Six Sigma concept.